The right question isn't "what can I automate with AI?" — almost anything can be. The question is "what pays back, and in what order?". Automating the wrong process first is the fastest way to burn budget and team trust. Here's the criterion we decide with: four filters and a matrix to score before you touch anything. The overview of automating with AI tells you what it is; this tells you where to start.
The 4 criteria that separate a good candidate from a waste of time
A process is a good candidate when it meets four conditions at once. Missing one doesn't rule it out — it drops in priority. Missing two? Don't even look yet.
| Criterion | The question you ask | Scores high if… |
|---|---|---|
| Volume | How many times a month does it happen? | It repeats dozens or hundreds of times a month |
| Repetition | Is it the same task every time? | The pattern is nearly identical each case |
| Clear rule | Can you describe the decision with explicit criteria? | A new hire could learn it from a one-page manual |
| Available data | Does the input exist in a machine-readable format? | The data already lives in an email, PDF, sheet or API |
The trap is the fourth criterion. Many projects die not because the process is bad, but because the input lives in someone's head, in a WhatsApp thread or on paper on a desk. No accessible data, no automation — just a digitization project in disguise.
The decision matrix: score every process before you touch anything
Take your candidate processes and score each criterion 0 to 3. Add them up. The total out of 12 tells you green, amber or red — and settles the opinion war in the meeting.
| Total out of 12 | Verdict | What to do |
|---|---|---|
| 10-12 | Green | Clear candidate. Into the first batch it goes. |
| 7-9 | Amber | Worth it, but fix the weak criterion first (usually the data). |
| 4-6 | Light red | Wait. The ROI lands late or needs too much supervision. |
| 0-3 | Red | Don't automate it. Redesign it or leave it to humans. |
The order of attack: what to automate FIRST
Among all the greens, don't start with the most ambitious. Start with the one that crosses two axes: high return and low implementation effort. That's the quadrant that gives you a visible win in weeks, funds the next project and convinces the team this works.
- High return / low effort — start here. Always. It's the quick win that buys credibility.
- High return / high effort — the second batch. Now that the team trusts it and you have muscle, go after the big one.
- Low return / low effort — slot it in as filler when there's room, not as a priority.
- Low return / high effort — don't touch it. Ever. This is where "AI for everything" projects die.
The classic mistake is starting with the flashiest thing — usually something customer-facing, complex and high-effort — because it impresses in the demo. Six weeks later, without a single delivered win, the project loses backing. Start small, ship, and use that proof to unlock the big stuff.
How three real processes score
The matrix lands better with examples. Here are three typical processes scored on the four criteria. Notice how the same business can have a green and a red side by side.
| Process | Volume | Repetition | Rule | Data | Total /12 |
|---|---|---|---|---|---|
| Answering support FAQs | 3 | 3 | 3 | 3 | 12 · Green |
| Classifying and routing incoming tickets | 3 | 3 | 2 | 3 | 11 · Green |
| Negotiating a custom contract | 1 | 0 | 0 | 2 | 3 · Red |
Answering FAQs scores 12: high volume, identical task, clear rule and the data already lives in your knowledge base — the archetype of customer support automation. Routing tickets loses a point on "rule" because of edge cases, but it's still green. Negotiating a contract scores 3: every case is different, there's no explicit rule and much of the judgment lives in the rep's head. That process doesn't get automated — at most it gets assisted, as covered in automating sales with AI.
The processes you should NOT automate (even if they score)
Some processes score green on the matrix and are still best left alone. Recognizing them saves money:
- The ones about to change. Automating a process you'll redesign in three months is wasted work. Stabilize first, then automate.
- The regulated-decision ones without clear human oversight (AI Act, GDPR). The data exists and the rule looks clear, but the risk isn't worth it.
- The ones that are your value proposition. If the human touch is what sets you apart, automating it is competing against yourself.
- The ones with deceptive volume. Lots of different emails look like one process, but if each needs its own judgment, there's no real repetition to capture.
From criterion to production
The matrix is the map; the work is driving. Scoring your processes takes an afternoon and saves you months of automating the wrong thing. What comes next — building the first green, leaving it running with an owner and metrics, and chaining the second — is where the craft is.
If you'd rather not build the matrix alone, operations automation does exactly this: we map your processes, score them with you, tell you which YES, which NO and in what order, and leave the first one in production. We don't hand over a report with recommendations — we hand over the process running.